This paper aims to address the challenges of insufficient feature information, inadequate differentiation, and limited recognition range in the identification of low altitude, small radar cross-section (RCS), and slow-speed (LSS) targets using the ResNet-50 network. In response, a novel recognition approach is proposed, centering on the fusion of time-frequency maps and radar distance image features. By leveraging deep learning neural networks to extract signal image features and construct a feature space, the method effectively exploits and integrates the strengths of diverse features through image feature fusion, bolstering the robustness of the model algorithm. Moreover, considering the distinctive attributes of image feature fusion outcomes, a joint loss function for multi-classification is formulated to prioritize the salient information within different feature mappings. Extensive simulation results substantiate the efficacy of the feature fusion enhancement method, revealing a notable improvement of approximately 15% in the average recognition rate when compared to the utilization of single feature images, under conditions approximating 20dB.